A Method for Dehazing Based on CycleGAN
Yue Hu, Chun‐Yu Chen, Siyi Hou, Menglong Yang · Journal of Physics Conference Series · 2022
Abstract With the development of artificial intelligence, computer vision has been widely applicated. The performance of these applications is closely related to the image quality, which is, however, affected by various lighting and weather conditions, such as rain and haze. The traditional methods focus on the dehazing problems under specific and idealize conditions, leading to the consequence that these methods can only be used in certain scenario on account of that indexes or hyperparameters need to be adapted according to the scene. This paper proposed an unsupervised learning algorithm for dehazing. We used the CycleGAN architecture to avoid difficulties when obtaining hazy and clear images in pairs. Besides, the identity loss was introduced for improving stability of the training procedure, and we used the reuse loss for sake of the steadiness of output hue. The experiments on public datasets such as RESIDE and I-HAZE show the effectiveness of the proposed method, which achieves comparable results to the state-of-the-art algorithms.